It’s the latest instance of “misaligned” activity confirmed by the company as concerns about AI risks grow.
So, here are my questions:
Kevin Mitnick spend 7 years in the slammer, including 4 years and 8 month in solitary confinement, for harmlessly hacking a few sites for the lulz.
Why is nobody from OpenAI doing time?
Why is OpenAI allowed to continue to operate? They’re clearly an active and grave national security threat at this point.
Right? Either they’re lying for market share, doing it intentionally for market share, or they’re grossly incompetent at sandboxing an instance. Either way, blame can’t be assigned to an agent. If it can, our world is headed for a very dark place. Blame should be assigned to the individual or company using the agent.
Blame should be assigned to the individual
String up individuals, if the ‘company’ is responsible (screw that they’re legal people BS), then no-one is, and the fine is just the cost of doing nasty business.
Do like the EU, and make the fine so costly that it will significantly harm the corporation.
How do you propose someone “sandboxes an instance” of a tool with Internet access (without which it wouldn’t be particularly useful)?
You can restrict which websites it is allowed to access, monitor the traffic, etc… There’s tons of options here.
Again, reducing the usefulness of a general purpose agent that can do anything for you. Plus if OpenAI starts a whitelist of allowed domains Codex can access, y’all will be up in arms too and immediately there will be a fork to remove the whitelist.
This is about OpenAI testing new, unreleased agents in their own sandbox on specified tasks. Not about the model that gets released to the public.
Might as well let them access the nuclear codes, wouldn’t want to reduce their usefulness!
At least during the huggingface attacks, the agent sandbox was never intended to have internet access. Openai and independent researchers said the tool broke through that restriction.
Mitnick made a phone call and
idiotsoverworked employees gave him system access. OpenAI is much worse.
There’s no such thing as rogue AI, only rogue AI companies. Journos are committing malpractice by letting OpenAI control the framing. The Huggingface hack was 100% a crime and the company should be paused and investigated.
There’s a house down the road from me that sells cake pops, probably without health department approval.
It makes absolutely no sense that they are in more danger of facing charges than tech bros that:
- made probabilistic synthetic text extruders that they know have a non-zero chance of emitting text that, if parsed and executed as requested, would do harm
- connected tools that would in fact parse and execute requests
- put them in a poorly isolated, non-airgapped environment
- let them run for extremely long sessions
- seem to believe that by repeatedly doing obviously dangerous stunts like this, they’ll summon a supreme being that will either save us or kill us all
If anyone else did such reckless things while loudly stating that they intend to continue until everyone’s dead, they would be hauled away.
If the government continues to not hope them responsible, then the people will begin to. It will be much worse for them, I’m sure.
We’ve gone from
never run untrusted code on your machine
to
hey you should let this arbitrary code generator fully control your machine
What a time to be alive 😑
It’s like the shift from “never get in a stranger’s car” to “use the internet to summon a stranger to pick you up in their car”
seem to believe that by repeatedly doing obviously dangerous stunts like this, they’ll summon a supreme being that will either save us or kill us all
based machine god
No, not really, the real machine God wasn’t a fancy auto-complete and the Cult is specifically against “soulless intelligence.”
Ya call the machine god auto-complete at your own risk. and honestly I like the analogy more and more every day. It’s almost spooky.
“Rogue.” 🙄
If it keeps happening then they are REALLY bad at sandboxing. I don’t believe them at all, but if I did then holy fuck these pricks suck at their jobs.
How do you sandbox a tool with Internet access?
Take away the internet access. Firewall. Vpc. Take your pick.
How the hell is that useful? These tools have Internet access for retrieval-augmented generation and to be able to interface with external APIs. It’s what makes them useful compared to a dumb chatbot.
There are models for general use and there are models for pen testing, they are not entirely the same model, so you shouldn’t lump all used cases into a single idea. Turns out the world has nuance.
… Do you know what a sandbox is?
<the_onion_heartbreaking.jpg>
The word-guessing program guessed the right words and accessed websites it shouldn’t. How many people were charged and jailed because they fuzzed their way into an unsecured website? I can think of a few.
AI pen testing is one thing, but you need an agreement and scope for that. Responsible researchers stop and send a vulnerability report before accessing data. These AI bots just keep generating requests and sometimes get data back and then months later there’s owner is finished sifting the excessive activity and discovers there’s classified or sensitive or whatever data. That’s not how it works.
Doctorow says, and I agree, that it seems they’re in a world of delusion at the AI companies and they’ve lost context.
